The US Pulls Fable 5, Open Source AI Manifesto, and Local Agents on macOS

Good morning. It's Saturday, June 13, and we're covering an export control directive that yanked Anthropic's most capable models, a manifesto for open source AI, and a guide to running local coding agents that do not depend on any cloud provider.

The Fable suspension thread hit 2,322 points. The open source AI manifesto hit 933. Both stories are about the same thing: who gets to control the models you build on.

In today's brief:

• US government orders Fable 5 and Mythos 5 suspended for all foreign nationals
• Ahmad Osman argues open source AI is existentially important
• A practical guide to local coding agents on macOS
• The architect-loop pattern cuts Fable token usage by 80%

THE BIG ONE

US Government Orders Fable 5 and Mythos 5 Suspended

The US government issued an export control directive requiring Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals, both inside and outside the United States, including foreign Anthropic employees. Anthropic complied by disabling both models for all users roughly four hours after the directive arrived at 5:21 PM ET on Thursday.

The stated trigger: a jailbreak demonstration. Anthropic's characterization: asking the model to read a specific codebase and fix software flaws, a capability the company says is available in GPT-5.5 without any bypass. The vulnerabilities surfaced in the demonstration were already known and minor. Anthropic's statement explicitly disagrees with the finding, arguing that "if this standard was applied across the industry, it would essentially halt all new model deployments for all frontier model providers."

The HN discussion and a critical analysis (250 points) raise harder questions: Is this principled regulation or political retaliation? Why does the directive uniquely burden Anthropic while leaving GPT-5.5 untouched? What does this do to the trillion-dollar infrastructure bets on closed frontier models?

Why it matters: The precedent is established. The US government will recall a deployed commercial model affecting hundreds of millions of users on the basis of an undisclosed verbal report with no statutory process. Every company building on closed frontier models just watched the off switch get used.

Our coverage: why the US government pulled Fable 5, the June 22 decision checklist, model dependency risk after Fable 5.

PLATFORMS

A Manifesto for Open Source AI

Ahmad Osman published "Open Source AI Must Win" (933 points), arguing that artificial intelligence represents critical civilizational infrastructure that should not be controlled by a small number of proprietary companies.

The core thesis: if intelligence becomes available only through paid access to closed systems, society loses more than software freedom. It forfeits operational independence. The ability to study, deploy, modify, and run AI systems locally is existentially important when models control work, education, science, and creativity. Reliance on closed APIs creates dependency risks that the Fable suspension just made concrete.

The timing of the manifesto landing on the same day as the Fable suspension is not coincidental. The argument gains force when the government can disable your model layer with a Friday evening phone call.

Why it matters: "Subscription economy for cognition" is the phrase that captures it. The alternative requires distributed, locally deployable AI infrastructure that does not ask permission from any vendor or government.

TOOLS

How to Set Up a Local Coding Agent on macOS

Kyle Legg published a practical guide (375 points) for running coding agents locally using llama.cpp with Metal acceleration, Gemma 4 26B-A4B at Q4 quantization, and the Pi terminal agent.

The key performance claim: Multi-Token Prediction pushes inference from 58 to 72 tokens per second on Apple Silicon, a 24% improvement over baseline. The guide covers building llama.cpp with Metal and Accelerate flags, downloading model files (roughly 17GB), running llama-server with MTP parameters, and configuring Pi to connect via the OpenAI-compatible API endpoint. Legg also notes that llama.cpp outperformed MLX on this setup (72.2 vs 45.8 tokens per second) despite MLX being Mac-optimized.

Why it matters: Local inference means no API costs, no rate limits, no vendor dependency, and no Friday surprises. The capability gap between local and cloud is narrowing.

TOOLS WORTH A LOOK

1. Architect-Loop - Cross-vendor agent coordination that reduces Fable token consumption by 80% through role specialization. Fable handles planning and review; Codex handles execution in isolated worktrees. Each builder runs in its own git worktree with explicit scope boundaries. (OSS, 94 points)

2. BitBoard - YC P25 analytics workspace for agents. Track agent runs, costs, and outcomes across your organization. (paid, 43 points)

3. Ezra - Lightweight task queue on Erlang/OTP, SQLite-backed. Designed for reliability and simplicity over distributed scale. (OSS, 38 points)

WHAT ELSE IS HAPPENING

• "There is a shadow hanging over this Fable thing" (250 points): Critical analysis of whether the directive reflects principled regulation or political targeting. (5 min read)
• "Slightly reducing the sloppiness of AI generated front end" (194 points): The simple fix is requesting Qt app aesthetics. Design system references beat abstract quality requests.
• OpenAI WebRTC Audio Session with document context: Simon Willison's updated tool for voice conversations with GPT-Realtime-2 now includes optional document context support.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.